Class 3: Artificial Intelligence in Finance

[SQUEAKING] [RUSTLING] [CLICKING] GARY GENSLER: I'm going
to talk a little bit about how I see financial
technology in a stack. We did a little bit of this
in our last two classes, but just really thinking about
the finance technology stack and then turn back into
AI and machine learning. In finance, we reviewed this
a bit in the last class, but I just want to go back to
it and talk a little bit more about it, a little bit more
granular this time around. And then take the bulk of the
class around public policy frameworks and how
AI fits into that, Artificial Intelligence,
machine learning, chat box, and the like. So that's sort of
the run of show. And Romain, you let me
know when we have either 15 or 10 minutes to go because
I don't have a clock here on this Zoom. So the three readings really
built upon last class's readings, but they
had a little bit of a tone towards
the regulatory side. And Oliver Wyman and
the Mayer Brown– Mayer Brown is a law firm. Oliver Wyman thinks about risk
management and consultant risk.

But each were with a
little bit of a gloss on how to manage if
you're at a board level. And the Oliver Wyman really went
through the lines of business, how machine learning and
artificial intelligence is being used. And the Mayer Brown went
through sort of the laws. And we'll come back to that. But that's why I sort of
reached out and had that. Now, the short 1- or 2-pager
from Julie Stack from the Federal Reserve, I thought
it was just interesting to say, all right, here's a senior– and Julie's very well-respected
in the community– a senior career person
at the US Federal Reserve writing about fintech.

What are they thinking? And just like I did earlier,
and we shared sometimes, what is the official sector
thinking, as we shared the chair of the
FDIC's speech earlier, and I will throughout
this semester, I think it's helpful
if you're thinking about this to sometimes
think, all right, what's the official sector
saying in their speeches and so forth? But that was kind of why I
put those readings out there. I'm looking at Romain to see if
there's anything, but probably not yet. And then the study questions. This is where it gets
a little bit more fun and I see if we can get
a little engagement. And again, we spoke a great
deal about this already.

But if anybody just really
quickly want to give an articulation
of this question– why are these new forms of
AI enabled data analytics, pattern recognition, speech
recognition, and so forth– how do they fit into the
other trends, as you see it, as we've talked about
in financial technology? GUEST SPEAKER: Luke? GARY GENSLER: This
is your fun time to either call on people
as they've volunteered or otherwise. But we've discussed this. I'm just trying to get
it a little going here. GUEST SPEAKER: We
can start with Luke. AUDIENCE: So the commonality
among the industries, so sector agnostically, is the
fact that all the companies who can deploy this AI to
their operation system is to save money, save costs, so
that the bottom line is better. GARY GENSLER: So one thing that
Luke's raising is saving costs. Others want to sort of chime in. And I'm kind of also
curious as others chime in how you see it fitting
in with other emerging trends. We had the trends that
we've already lived through, but we're building upon,
like the internet and mobile and cloud.

We have some other
trends that we'll get to in future
classes like open API. Just kind of curious
to see if somebody wants to take a
crack at connecting this piece of the
technology with some of these other trends. GUEST SPEAKER: Laira? AUDIENCE: Yeah, I think what
we discussed extensively in the last lecture
was Erica being one of really good examples
of how the new forms of AI is emerging with
what's already existing and making it not just
cheaper for the firms to answer mundane questions
that customers have, but also making it
more user friendly. So I think in terms
of Erica, it's just a great example to show
how this question kind of goes through. GARY GENSLER: So
Laira's just raising– and sometimes I just break
these down simplistically. In the artificial
intelligence world, there's the
consumer-customer interface. Erica at Bank America
is an example, and chat bots, and
the various ways we communicate with customers,
tying into customers, builds what technology? What is it that you
use and might even be using right now when
you're watching this course? GUEST SPEAKER: Eric, you
also had your hand up.

AUDIENCE: I was going to
talk about something else. I was going to say that– GARY GENSLER: Sure. [INAUDIBLE] and I'll
tie them all together. Don't worry. I'll answer [INAUDIBLE]
questions, too. AUDIENCE: Sure. I was saying that AI is being
used by fintechs for better underwriting
purposes, like using alternative data to better
assess people's credit. GARY GENSLER: Absolutely. So it's the data analytics.

It's the customer interface. That data analytics of
predictive underwriting, whether it's in
insurance, whether it's in lending, predictive
underwriting. It's also on the
customer side where we use natural
language processing and we interface
with the customers. Romain why don't we
take one or two more? I'm looking for
somebody who wants to tie it to the other
technological trends that we see. GUEST SPEAKER: So let's go
with Nikhil and then with Wei. GARY GENSLER: All right,
and then we'll move on.

AUDIENCE: I think the
Oliver Wyman reading talks about how companies that have
been using AI and machine learning have done better. I think it was
asset management was a specific example they took. I think it also ties to,
like another class I'm taking with Simon
Johnson on AI talks about David Autor's report
that says there's a superstar effect where firms that
have access to this data and are using AI tend to
perform better in the market. And I think that's a
significant tie-in. And it's probably
even more exaggerated in fintech specifically. GARY GENSLER: So let's
just pause for a second. It's data, data. What we've had is this
remarkable advancement in data analytic tools,
artificial intelligence. But we've also had a remarkable
advancement of the ability to store and process
data through the cloud and just through the emergence
of much faster computers and much more connected
communications. So that data piece, the
artificial intelligence and machine learning
trend might not have been able to do as well
if it weren't for everything that's going on,
broadly speaking, whether it's in
cloud or computing.

And Romain the last
person that was– AUDIENCE: Yep. So I also want to maybe make
a mention that in also help was a lot of times either
collecting data analytics or cleaning the data analytics. Because a lot of time
that in the old world there's a lot of data
you potentially collect. First of all, I can
help to better collect unstructured data. And the second found
that it helps to clean a lot of data you collected. GARY GENSLER: So
absolutely agreed. And often it's 80%, sometimes
90% more of a computer science group is in the cleaning up of
data and standardizing data.

And we'll come back
to this, but a lot of fintech disruptors,
a lot of startups have actually created value more
around data than anything else. And I will say not just about
data, but standardizing data. And later in this class, we're
going to talk about Plaid and Credit Karma, both of
which were earlier this year acquired, Plaid by
Visa, Credit Karma by Intuit for $5 to $7
billion– big, big acquisitions.

And we're going to talk about
what was the value proposition for Visa and Intuit? Why were they paying
$5 or $7 billion? A lot of it– not all of it, but a lot
of it relates to data, but also having standardized
that data, particularly in the case of Plaid. How it's affecting the
competitive landscape. We've talked a great deal. Hopefully this
will continue to be a theme throughout the semester
about big incumbents, big tech, and fintech startups. I will contend in this
and throughout this course that AI and machine learning is
now moving into the technology stack. If we think of this stack
as layers of technology that incumbents incorporate,
and frankly will not survive if they
don't incorporate, that AI and machine learning
is being incorporated quickly into the financial
incumbent technology stack.

We're not fully there yet. And that the competitive
landscape is such that the fintech startups
and disruptors have been able to find cracks
in the old business models. And using machine learning,
they've been able to break in. And we'll talk a bit about that. The big tech firms,
of course we've already talked about that they
are really about networks. Networks that they then
layer more activities upon, and those more activities
bring them more data. And then we're going
to talk a fair amount about public policy. But anybody who's sort of dug
into the Mayer Brown reading want to just give two or
three thoughts on the broad– what are the– I'll later call
them the big three? But it's almost written right
in the question for you, but Romain, you want
to call anybody? GUEST SPEAKER: Michael? AUDIENCE: Yeah, so
the reading kind of did touch upon bias a lot
and its potential, just on the natural factors
that a machine learning algorithm would trace.

GARY GENSLER: So one of the
things about machine learning and deep learning is
that it's remarkably successful at
extracting correlations. Correlations from data sometimes
that we didn't see before, that didn't come just from
a linear relationship, a linear relationship that we
might be able to identify just in classical statistics. But in those
remarkable abilities to extract correlations,
you might see biases. If the data itself has
a bias in it that people of certain gender, certain race,
certain ethnic backgrounds, certain geographies are more
likely to, in the data's mind– in the data's mind– are more likely to
have lower income and in the data might have more
likely to be a lower credit quality, then you
might be embedding certain biases inside the data. And many nations around
the globe, not just the US, have said to the
credit card companies and the other financial firms
that you shouldn't have biases around race, gender, ethnic
background, geography, sometimes, and the like. So one is biases. When I consider the
three big buckets here– anybody want to just
talk about the other two? Romain? GUEST SPEAKER: Alicia.

AUDIENCE: Hi. I think we talked
this last class. I think AI derives
conclusions or correlations without explaining the why. So humans cannot understand why
some guy has a better credit rating than another and has an
issue with the law, basically. GARY GENSLER: Yeah. And why as societies have
we embedded in laws– and we'll talk about this. But if you have a
point of view, why as societies have we
embedded in laws that you need to be able
to explain the why when you deny somebody credit or deny
somebody a financial product? We did this in the United
States 50 years ago in something called the
Fair Credit Reporting Act.

Data analytics was a big wave
in the 1960s, believe it or not, when credit cards were
invented in the 1940s and '50s. By the 1960s, data
analytics were going, and the Fair Isaac Company,
which became FICO, had started. And we embedded in law that you
had to answer this question. Explain why you denied credit. But why do you think we
embed that in country after country in our laws? GUEST SPEAKER: Danielle? AUDIENCE: So I think it's,
going back to the bias question, to prevent bias in people
who are extending credit. GARY GENSLER: I
think you're right. I don't think it's
the only reason, but I think it's
a dominant reason. We also in the US passed
something called the Equal Credit Opportunity
Act, or it generally goes by the terms ECOA.

But those two laws and
another law in the US, Truth in Lending Act
for transparency, were kind of this bedrock out of
the 1960s data analytic credit card boom. By the early '70s,
we had those three. Anti-bias, fairness, you
might say, explainability. These are two bedrocks
in finance in Europe and the US, country
after country. What's the third
challenge that comes up with data analytics or AI
that often we find ourselves, and if you're starting
a fintech startup you have to be aware of? Romain, any hands? GUEST SPEAKER: Not yet. We have Luke again.

GARY GENSLER: We'll pass on
Luke unless somebody else. GUEST SPEAKER: We
have Danielle again. GARY GENSLER: All
right, either one, whoever's got their mic off. AUDIENCE: So privacy
is the last one. GARY GENSLER: Sure. AUDIENCE: For example, companies
have demonstrated the ability to predict when consumers
have certain health conditions or pregnancy, for example. There is a really famous
case where a company knew that a consumer was pregnant
based on how their shopping patterns changed,
and there are reasons we've precluded
employers or credit extenders from asking about
certain parts of people's lives.

But we may be unexpectedly
exposed to parts of those lives if we're capturing
data and using it. GARY GENSLER: So this
trade-off of privacy versus financial
services, thought it's not as old as
sort of the fairness and the explainability,
which in the US and then later in
other countries was embedded in many
laws 30 to 50 years ago, privacy has picked up a
little bit more of a stream. By the late 1990s
in the US, there was modest financial
privacy protections that were embedded into law in 1999.

I actually helped work on
that with then-Congressman Ed Markey, now Senator Ed
Markey of Massachusetts. But in Europe, they
went quite a bit further in something called
the GDPR, which we'll talk about a little later. But the General Directive– P doesn't stand for
privacy, but I think it's Protection of Regulation. So those three buckets– those three buckets
are the important ones. So again, AI machine learning
fits into these other trends that we think about. And I'm going to
walk through that in this class of cloud and
internet and mobile and data. Fintech startups, big tech,
and incumbents, I believe, are all embedding it in
their technology stack. And you're really
challenged if you don't. And then the big three
challenges in public policy, explainability,
bias, and privacy. There are other
challenges as well, but those are the big
three, in a sense. So what do I mean
by technology stack? Well, I think that
three things are already embedded, the internet,
mobile, and the cloud.

And if this class were being
taught at MIT in the 1980s, none of them would be
there, and by the 1990s, we would have said,
wow, that internet. The word "fintech" didn't
really come about in the 1990s. But if we had applied
it to the 1990s, the internet was
dramatically changing. Mobile into the naughts
in the cloud and so forth. I would contend you cannot
really survive in the finance space giving customers
what they need, whether it's in the wholesale
markets of capital markets and payments, or in the retail
markets if you haven't yet embedded in your
technology stack.

Now, I will note that many
large financial companies are slow to use the cloud. The largest amongst them
tend to want to still have their own data centers. I think you're going to
see that shift dramatically in the 2020s. But I'm certainly telling you
that if you start a startup, you cannot survive if you're
trying to do your own data center, if you're going
to already embed these in your what I'll
call financial stack. The internet for connectivity,
mobile in a sense for ubiquity, meaning that folks
can be out there. Cloud, you're sort of renting
somebody else's storage and often their software.

But then the things that we're
talking about in this time, in the 2020s that
are being embedded into the classic standard
company stack is AI, machine learning, and natural
language processing, and what we'll talk about
in the next class a lot about open API. Now, we're in a transition mode. Not every company has really
embedded it in their stack. And these are where the
opportunities really existed in the last
dozen years in fintech.

Fintech startups that
were savvy enough to really build this into
their product offerings faster than the
incumbents, or, better yet, in a more refined, targeted way. And we'll talk a fair
amount about that. Now, of course, there's
other things in the stack. And this is not
what this class is. Even money itself in accounting
and ledgers and joint stock companies were all in a sense. We just take them
completely for granted. By the time you're in a
master's program at MIT, master's of finance or MBA
or other graduate program, you're quite familiar,
and you almost just take these for granted.

But I can assure you
at earlier decades, they couldn't be
taken for granted. And some of them, like
securitization and derivatives, will dramatically
shift your ability if you're doing a
startup to compete. I see some things in the chat. Romain, are there questions? GUEST SPEAKER: All good, Gary. GARY GENSLER: All right. And then the question
I sort of still have, and I teach this
quite a bit at MIT, is blockchain technology. Will that move into the stack? I would contend it's
not really yet there. You can be an incumbent. You can be a big finance firm, a
big tech, or a startup and say, I'm not going to
compete right there. I'm not quite sure. Though, again, we
look at Facebook. We look at Telegram, big tech
companies, messaging companies, [INAUDIBLE] in
Korea who are sort of pulling in some blockchain
technology and looking at it. We see trade
finance consortiums. And we'll talk more
about this next week.

But I would say
that you will not survive if you're not
bringing machine learning into your technology stack. You probably won't
survive that long if you don't really have
a strategy around open API and data. Romain, I pause a little bit. We talked last session about
artificial intelligence and machine learning. We're not going to dive back in. I'm just going to
open it if there's any questions about
what we talked about. That, of course,
machine learning is just a part of
artificial intelligence.

You narrow it down
to deep learning. Fundamentally as
a business school, I'm not asking each of you to be
able to program with TensorFlow and run a TensorFlow project,
even though many of you know how to. I'm sort of just saying to think
about, from a business side, it's about extracting from
data, cleaning up that data, standardizing that data,
and often labeling it.

Labeling it because
you can learn faster. That's called
structured learning rather than unstructured. But labeling that data
and then extracting correlations and decision
algorithms that come out of it. Romain, any? GUEST SPEAKER: Luke has
raised his hand again. GARY GENSLER: I'm
going to pause. AUDIENCE: Just a quick question. GARY GENSLER: Oh, a question. Yeah. AUDIENCE: Yeah, a question. So how can a country
that is developing fintech out of
not because it was underbanked, but
rather overbanked, but looking for
alternative investment– so the likes of South Korea– develop a bunch
of coders or those with– actually, better
yet, those people who can draw a conclusion and
extract hypotheses and build up better ways to
build an open API, how can a government
really step in to encourage that and
make an ecosystem? Somebody's got to do something.

And I'm not sure
America have a bunch of great coders and great
minds, and it's a melting pot. So [INAUDIBLE] bunch
of geniuses here. GARY GENSLER: Yeah, I'm not
sure I follow the question, but I'm going to take
it and then move on. I think what you're
saying is in a country that has a very
advanced banking system, how can a government
encourage this? You do it through
the education system.

You do it through, just
as we do in the US, promoting STEM education
and programs like at MIT. I think over time,
there is a challenge of how you adjust
laws and regulations. Finance is a highly regulated
space in every country. We're dealing with trust. We're dealing with
people's money. We're dealing with inherent
conflicts of interests that you can't escape
in the world of finance. And so trying to deal
with that with regulation but how we adjust with these
new tools that are in place. But that would be
how I'd answer it. So let's go back to what we
talked about and just sort of do a little bit more granular. I contend at its core that these
technologies, machine learning and natural language
processing and AI, need to be brought
into the finance stack and the technology stack. And so every type of
company, whether you're in payments or
you're in lending, whether you're
insurance or not, you want to think about how you
bring it in, whether you're a disruptor or not.

And so that's why
I think about it down the line in each of
these fields and not just about disruptors. And we talked about each
of these in the past, but I pause just, again,
it's a little repetitive. Just if there's any questions
about some of these slices. And then remember we're
going to be digging quite a bit into this
in the next five weeks as well in each of these areas. Romain? GUEST SPEAKER: I
don't see anyone yet. GARY GENSLER: All right. Well, what I've said is,
all right, so AI is a tool. And this is really
an interesting debate that people can have at MIT. And I've been in rooms
full of five to 10 faculty sometimes, sometimes one
on one, where we debate. Is AI a service,
or is AI a tool? And I would say that it's
an interesting debate.

Most of the time we
land on that it's more a tool than a service. But every new technology,
every new technology that comes along, whether it
was the telephone, whether it was the railroads,
whether it was airplanes, every new technology
that comes along has some attributes of being
a new industry, a new tool, the railroad industry,
for instance. And yet many businesses use the
railroad to do their shipping. AI and machine learning
is more like a tool than it is a service. But it doesn't mean
it's always just a tool. I did some research over the
last few days, just a list where we could go
through any one of these.

AI as a service. Here I've listed
10 or 12 companies in finance that are actually
doing AI sort of as a service. AlphaSense was started in 2011
before the whole rage of AI, but they were data analytics
as a search engine for finance firms to search key
terms and key words in registration statements
and other statements that are filed with these various
securities regulators around the globe.

Sort of think of it as the
Google for financial documents. Well, Google certainly has moved
dramatically into AI space. AlphaSense did as well. There's a number of these
in the insurance sector who really are around taking
photographs of automobiles at an accident scene, and then
based upon those automobile photographs or accident data,
to use machine learning. And so Cape Analytics,
Tractable are both firms that are in essence
providing services to insurance companies.

They have not yet, as to the
best of my knowledge, Cape Analytics or Tractable,
decided to have direct consumer interface. They're not selling insurance. They're selling a
software analytics tool to insurance companies. And similarly, like
ComplyAdvantage in the money laundering space
or Featurespace in anti-fraud. They're saying, we can build
something for fraud detection. We can build something
for this world of anti-money-laundering
compliance. We can build the software,
and we'll put our product out there for the banking sector
to basically rent us rather than building their own system. And you see others, document
processing and the like. And even Zest AI– Zest AI, founded in 2009,
before this conceptual framework and the big movement, but
Zest AI in credit underwriting software, basically providing– broadly speaking, I'm calling
it AI in finance as a service, rather than building it
right into the stack.

Romain I'm going
to pause for a bit. GUEST SPEAKER: If you
have any questions, please rate the
blue little hand, as you probably know by now. I don't see anything, Gary. GARY GENSLER: And I'd say
this, that each of these go back to some of
the sectors back here. So asset management, we
have sentiment analysis. We have– I'm not going to
pronounce the company's name right, but Dataminr. Dataminr, which can actually
do market sentiment analysis. And if you're a
hedge fund, you might sort of rent into Dataminr and
get that sentiment analysis. You have several
services that are doing fraud detection
and regulatory, the anti-money-laundering
slices. Credit and insurance– I picked three that are
doing insurance underwriting. But if you're Bank
of America or J.P, Morgan or Cap One in the
credit card business, you're going to be embedding
this right into your business, by and large. Not always, not always,
but by and large, you're embedding it
right into your business.

GUEST SPEAKER: We have
a question from Geetha. GARY GENSLER: Please. AUDIENCE: Hey, Gary. Geetha here. I work for our Capital One. I'm in the credit lending space. One– GARY GENSLER: You're
going to correct anything I say about Cap One. Please. AUDIENCE: No. One thing that I find really
surprising with the regulations is that if we develop our
own AI models, regulations– anybody, like when
auditing happens, they are very specific
about explainability, interpretability. But if you were to use a
vendor, Shape Security, Akamai, they don't care too much
about explainability. That I always found surprising. Why is it that within
the realms of bank, they're so specific
about regulations, but when we use a vendor,
the extent to which they care as how you use Akamai. You use Shape, and that's it. GARY GENSLER: Gita, I'm so
glad that you raised this.

I think I earlier had said
in our introductory class that one of the competitive
advantages of disruptors is that they have
certain asymmetries. Incumbents, like
Cap One– and if I might say you
working for Cap One, which is one of the big
seven firms in credit cards, is truly one of the incumbents. Incumbents tend to need to
protect their business models. And part of what
they're protecting is also the reputational
and regulatory risks. But the disruptors have a little
bit of a different perspective. They're not generally protecting
any in inherent or incumbent business model.

And yes, they're
also willing to take more risk with the regulators. I'm not saying whether
they should or shouldn't. I'm just saying this is kind
of the facts on the ground that disruptors are
a little bit more towards the end of basically
begging for forgiveness rather than asking
for permission, if you sort of
remember how you might have been with your parents or
if any of you have children. And incumbents are
more into asking for permission of your at
least internal GC, your General Counsel. Can we do this? Can we do that? And so I think the vendors– in that case, the vendors
are a little bit more willing to take risks
when explainability and the explainability
that's inherent in US law and the Fair Credit
Reporting Act and the like. And it doesn't mean that it's
more legal for a disruptor to do it than for
Cap One to do it.

It's just their
business model tends to be a little
bit more accepting of that regulatory
compliance risk. Secondly, and I think
this is probably a bit of a misreading
of the risks, but sometimes the thought is
if the vendor does something, it sort of insulates
the big firm. Now, my understanding–
again, I'm not a lawyer– but my
understanding, it doesn't really insulate Cap
One, or it doesn't really insulate Bank of America if
their vendor does something that's blatantly a violation. I don't think it does. But sometimes there is a
bit of that mindset as well.

Does that help? AUDIENCE: Yeah. Yeah. And the last thing is– not
taking too much time, just a comment. One other thing I find very
intriguing with vendors is that they often get
the data of incumbents, like maybe Bank of
America, Capital One, and they charge us
back for that data. That's the other thing
I find [INAUDIBLE].. GARY GENSLER: So here,
this is about how companies can capture data. We're going to talk a lot
about this one, open API. Just to the intersection,
this is one of the key things to take from these
series of classes. Machine learning is nowhere
if it doesn't have data. Data is facilitated by a lot
of startups getting access to that which the
incumbents already have. So around the globe, in
Europe, in the UK, in Brazil, in Canada, US, Australia,
there are significant efforts to promote what's
called open API– Application Program Interface. In essence, that is you or
I permissioning a company to get my or your data from
an incumbent financial firm.

And so we permission somebody
to get data, in Gita's example, from Cap One. Then they use that
data, the startup. And then Gita's saying
that Cap One then has to pay a fee
to some startup, even though the data had
initially come from Cap One. And that's a perfect
set up to two examples I just want to talk about. I want to talk about
two large mergers that were announced in 2020. The first one is Credit Karma. Now, I don't know if we
could do by a show of hands, but how many– just raising
the blue hands in the windows– how many people, if you can
go into participant buttons, have actually used Credit
Karma, that you would consider yourself one of these members? And Romain you'll tell
me what it looks like.

GUEST SPEAKER: We have at
least 10 students so far. GARY GENSLER: No, but
if you scroll down. So all right, so it's not as
big a percent as I thought. Let me go back. So Credit Karma started in 2007. The entrepreneur who started
it couldn't get a free credit report. So they say, why don't I start
a credit report platform? 13 years later, they were
able to sell for $7 billion to Intuit. Now, you might not be
familiar with Intuit. Their main products at Intuit
are tax software, TurboTax. They also have something
called Quicken Books. And I believe it's possible
they have a third product. They might even have Mint.

But Intuit saw they wanted to
buy Credit Karma that had never gone public. Credit Karma apparently had
nearly a billion dollars in revenue last year,
and yet Credit Karma is still a free app. How is it that
something that doesn't charge anything can have a
billion dollars in revenue? It's that they're
commercializing data. And remarkably, 106
million members– 106 million members. 8 billion daily decisions,
credit decisions or other analytic
decisions that they have. And so Intuit is saying, why
are they buying Credit Karma? Even at seven times revenue,
that's a healthy price. They're buying it largely
around data and data analytics. And credit card has figured out
how to basically commercialize that flow of data on over
100 million accounts. And how do they do that? They do it by cross marketing. So they're marketing
not just to us. But then they're also,
with many financial firms, they're going back and
say, this account here, this is a worthy thing.

So they make arrangements. They enter into
contractual arrangements with financial institutions
and then market to us to take a mortgage,
to take an auto loan, to take a personal loan. Plaid. Plaid's a company that
we'll talk a lot about when we talk about open API. This was software that
started just seven years ago. Two developers, straight kind
of hard-core computer scientists who had went to work for Bain. And for anybody who's thinking
about working for a consulting firm, this is not a vote
against Bain or BCG or others. But they basically decided
after a year at Bain to go out and do
their own startup.

And it was a startup to
facilitate financial disruptors or fintech companies
accessing data at banks. And there was not a standard. There was not a standard
for this open API. So they created at a hackathon– at a hackathon that
they actually won back, I think, in 2012 or 2013– they were in their late 20s
at the time, by the way, if you're just
trying to figure out. Seven years later, in
their mid to late 30s, they sell their business
for $5.3 billion. But it all starts at Bain,
computer scientists creating open API software. Well, what happened
over those seven years, 11,000 financial
companies signed up to use that standard
protocol to do open API. And all the other side,
2,600 fintech developers. And if anyone here has taken
Michael Cusumano's class on platforms, this is
the classic sweet spot of creating a
platform company, when you have this two-sided
many-to-many market. Many fintech developers want
to access financial data at financial firms.

Many financial firms don't
want to deal with thousands of fintech developers. And so inside of this
many-to-many market, Plaid creates a software,
a standard software for that to happen. But what did they
build on top of that? They built data aggregation. They announced a $5.3
billion merger to Visa. There's a lot of
people that debate whether it was a good idea
because the estimate by Forbes is there was only about
$110 million of revenue. I mean, now we're talking
40 or 50 times revenue. But 200 million
accounts are linked. We'll chat about this
more because those 11,000 financial firms could
all stop using Plaid and go to one of
Plaid's competitors now that Plaid's bought by Visa.

But this gives you the sense
of the power and the value, the economic value of data,
machine learning, and the like. Romain questions? GUEST SPEAKER: None so far. It seems like the
class is quiet today. GARY GENSLER: All right. So I'm going to talk a little
bit about financial policy. How does this all fit in,
in the next half hour. Broadly, first, is
just a sense of– I'm trying to get rid of it
this participant window here for a minute.

So broad public
policy frameworks have been around for thousands
of years, since the Hammurabi code, since Roman and
Greek times, sometimes embedded even in religious law. That's the nature of money. But four slip
streams, and all four will be relevant for
fintech as we talk through not just AI, but all sectors. One is money and lending. We've, over centuries,
often get official sector as a point of view,
sometimes even limiting interest
rates and the like. Two is financial stability. We think about a crisis.

We're living through this
corona crisis right now. Central banks around
the globe are, with an eye towards
promoting the economy, also thinking about how to
ensure for financial stability. The reverse of
financial stability was happening in 2008 crisis,
that that crisis where banks were faltering and closing up. And then that led to millions
of people losing their jobs, millions of people losing
their homes, and the like.

So financial stability,
I grab a couple pictures here out of the
Great Depression, an earlier period of crisis. But what we'll talk a lot about
is the third and fourth bucket. The third bucket of protecting
consumers and investors. Consumer protection we can
think of even just in terms of ensuring that if we buy
a crib for our children that it's safe. If we buy a car that it
actually is safe on the road. So consumer protection
refers to things much broader than finance. Investor protection is
the concept that, yes, we can take risk in markets. We're all allowed to
take risk in markets. But that the markets themselves
and the issuers, the people raising money, should explain
to us at least the material pieces of information upon
which we would take those risks.

And that the markets themselves
have a certain transparency and we protect against fraud
and manipulation and the like. And then guarding
against illicit activity. This is one that
we've really layered over the financial sector
in the last 40-odd years. In an earlier era, 19th
century, earlier 20th century, we didn't have as much about
this, even though, of course, we did guard against
bank robbers. But I'm talking about
illicit activity outside of the financial sector– money laundering, terrorism,
and even sanctions. So these four slip [streams
in a sense, are there. So how does it fit back to
AI and policy in finance? So I've talked about
what I have come to call the big three, biases,
fairness, and inclusion; explainability; and privacy. And what we mean by that
is if you take a whole data set, millions or tens of
millions of pieces of data, and extract correlations,
and you find patterns, some of those patterns
might have biases.

And those biases
can exist because we as a society are not perfect. We have biases even in
what we've already done. And so now you're
extracting and you might be embedding some of those biases. Secondly, sometimes
it will happen just out of how you
build your protocols, how you build your
actual questions and query on the data. But I assure you that most data
sets have some biases in them.

You just might not
be aware of them. And even if you have
a perfect data set, the protocols themselves might
sort of build some biases on top of them. And we're finding this in AI
policy not just in finance. It's true in the
criminal justice system. It's true in hiring, that
using machine learning, you have to sort of say,
wait, is there bias? And the laws here in the US that
are most relevant in finance started with something called
the Equal Credit Opportunity Act. And we'll talk a little
bit more about that. Explainability and transparency,
as we talked about earlier, is sort of like a cousin or
a sister to the bias issue.

And in the US, it
was 50 years ago that we passed these twin
laws within four years. The second law was the
Fair Credit Reporting Act. And this was the concept
about holding that data, but also being
able to explain it. Romain I see the chat button. GUEST SPEAKER: We have
no question from Jorge. GARY GENSLER: Please. AUDIENCE: Yes, professor,
thank you so much. I just want to have a
little bit more color on financial inclusion,
and specifically on what type of data,
what models are used? What's the forefront
of data modeling for using AI and to help
financial inclusion? Thank you. GARY GENSLER: I'm not sure,
Jorje, I follow the question.

Let me see if I do it, but
please keep your audio on so we can engage here. Biases are sort of the
reverse of inclusion. So financial
inclusion is a concept that everyone in
society has fair access and open, equal access in some
way to the extension of credit, to insurance, to
financial advice, to investment products
and savings products that they wish to, or
payment products as well. And the reverse of
inclusion is sometimes that somebody is excluded.

And excluding someone
could be excluding them on something that is allowed. Like I might exclude
somebody only earning $50,000 a year from an
investment product which is for
high-risk investors, depending upon how the
country is arranged. But in the US, we have a
little bit of this concept that sophisticated
investors can be investing in products of higher risk. Or at least they
get less disclosure. But that's how inclusion
and bias are kind of the– they complement each other.

The greater inclusion you have– you can get to greater inclusion
if you have fewer biases, in a [sense fairness. But [INAUDIBLE]. AUDIENCE: No, I was just– I totally get that. I was just trying to
understand what type of models, what type of data, what is
the forefront of AI currently? Because I totally get
it's gathering data and finding patterns. But digging a little bit
more on that, what type of– GARY GENSLER: So let's
say that one pattern that we know about already– this is a classic pattern
in credit extension. I don't know how many of you
know what retreading a tire is.

A retread means that you're
putting rubber on your tire. Instead of replacing your
tire on your automobile, you're actually paying to
put new rubber on the tire– tire retreading. It's been known for decades that
people who retread their tires are a little lower income,
generally speaking. And actually there's research– I don't mean academic
research, but there is research in the
credit business that retreading tires
means that you're probably a little higher credit risk. Now bring it forward to now. Bring it forward to
the 2020 environment. And let's say that
you can follow that those people who
bought tire retreading, or even if you went to
a website on your laptop about tire retreading,
let's say that's built in to an algorithm
that's going to give you lower extension of credit. That might be allowed, or it
might embed a different bias. It might be that tire retreading
shops are perfectly acceptable in certain communities,
either ethnic communities, or gender-based, or
racial communities, that it's just perfectly– it's not about creditworthiness.

So it's how you extract
certain patterns that are about credit extension
but not about race, ethnicity, cultural backgrounds,
and the like. And if you hold
just for a second, we're going to talk a
little bit more about this because I'm going to talk about
the Equal Credit Opportunity Act. AUDIENCE: Thank you. GARY GENSLER: Romain we good? GUEST SPEAKER: All good, Gary. GARY GENSLER: So beyond what I'm
sort of calling the big three, I list four other things. But they're all really relevant. They're relevant to, broadly
speaking, the official sector. But they're also
relevant as you think about going into
these businesses the use of alternative data.

And we'll come back to that. Basically, we've had data
analytics in consumer finance since the 1960s. We have in 30-plus countries
used these FICO scores. But beyond what is built
into the classic data set, what about new data? We have issues about whether
the algorithms themselves will be correlated or even collude. And this is absolutely the
case that one machine learning algorithm and another
machine learning algorithm can actually train
against each other. We've already seen this
in high-frequency trading. Even if the humans
aren't talking, the machines will
start to actually have a sense of cooperation. And when is
cooperation collusion? When is it that they're
spoofing each other or doing something
against each other in a high-frequency world? I deeply– and this
is one of the areas I want to research
more with colleagues. I deeply am concerned
that a future crisis– it's remarkable. We're in the middle
of the corona crisis, but a future crisis we'll
find algorithmic correlation.

And this is certainly the
case in smaller developing countries, where a Baidu from
China or a Google from the US might come in, and they might
come in with their approach to artificial intelligence. Or a large financial firm– it could be a European, Asian,
or US financial firm comes into that smaller
country, and they kind of dominate the thinking about
how to do underwriting, and they're the
big network effect. And all of a sudden,
the crisis of 2037 might be that
everybody is extending credit kind of consistently
in the same way. So it's basically
less resilient. We're living through
a moment of crisis right now where we're testing
the resiliency of humankind through the corona crisis. But I'm talking one
in the financial side. And then the question is,
how does machine learning fit into current
regulatory frameworks? Around the globe, a
lot's been written, but it's all at a very high
level, and it's non-binding. But it's these top
three, as I've mentioned.

So the alternatives– this
was a question earlier– is the official sector can
stay neutral and say, listen, this is just a tool. We're still going
to regulate lending. We're going to regulate
capital markets. We're going to regulate
everything the way we did. And new activities will just
come into those frameworks. Maybe we'll clarify a little
bit around the fringes. Secondly, you can adjust. Turn the dial. When the internet came
along in the 1990s, at first it was like
technology neutral. And then pretty much every
regulator around the globe had to adjust. What did it mean if there
was an online bulletin board that was trading stocks,
where buyers and sellers can meet? Was that what's
called an exchange? Should it be regulated
like the New York Stock Exchange or London
Stock Exchange, or regulated maybe a
little differently? Where our securities regulator,
where the Europeans ended up in the 1990s was to regulate
these online platforms like exchanges
but not identical.

So then we had a regime of
fully regulated exchanges and these online electronic
trading platforms. And that was then later
adopted in Asia as well, with some variations. The other thing is that
the official sector often tries to promote this–
promote the innovations, promote the technologies,
or promote open banking, as I've said. But an interesting piece
of this all is activities. Should we think about
machine learning as a tool, like a hammer that
everybody is going to be using, like electricity,
like the telephone that everybody is using? Or should we think about it, as
I said earlier, some companies are providing AI as a service? Activities, a
technology, a tool. And the official sector
grapples with this sometimes. To date, mostly they've
stayed technology neutral with a little
bit of promoting early stage activity and the
promoting of the open banking. Romain questions? GUEST SPEAKER: We
have 15 minutes left. GARY GENSLER: OK. Alternative data.

This is data that
you can extract, whether it's banking and
checking information, or as Alibaba does in China,
taking a whole cash flow approach. Saying I can see everything
about your business. It's called cash
flow underwriting. Here in the US, a
payment company, Toast, was able to do– until
restaurants closed down– able to cash flow underwriting
around their restaurants because they had
that payment data. All the way down to your app
usage and browsing history. I believe this is
a trend that we've been on that will be accelerated
by the corona crisis. That this crisis, we're
finding whether it's the large firms like Google
or even smaller ones, want to contribute to trying to
thwart the virus by following us or our location devices. Our location devices, of
course, are called cell phones and smartphones. But such vast parts
of the population have them that with
location tracking, we can possibly thwart or even
contain this virus by watching how we track ourselves.

I think that we will
shift a little bit further into data sharing, and
that will look back and 2020 will maybe be,
and 2021, a pivot point. But what it does
mean, even in finance, that a lot of this data is
going to be available somewhere, even more, maybe, than
currently is available. So there are actually
alternative data fintech companies. CB Insights, which is a
leader in tracking fintech, puts together this chart. And we don't have time to
go through these companies, but these are companies that
are sort of marketing themselves in this alternative
data set, almost like capturing the data, and
then data and AI as a service.

I want to just talk about
Apple Credit Card for a second because it's where you
can also stub your toe. Apple Credit Card with a big
rollout in conjunction with Goldman Sachs' Marcus
and MasterCard– so it's a really interesting
combination of big tech, big finance together– rolls out a credit card product. I think it was in November– yeah, November of
this past year. And in rolling it out
in a very proud rollout, an entrepreneur, here going by
the Twitter account DHH, which you might know of
this, goes on and finds that he is provided greater
credit than his wife, and he and his wife
are both billionaires, like literally worth
a lot of money. Maybe it was only they were
centimillionaires, but worth a lot of money. Joint tax account,
joint assets, and he was being provided 10
to 20 times more credit. So he took to the Twittersphere
and sort of made this. And he says his wife had
spoke to two Apple reps, both very nice. But basically saying,
I don't know why.

It's just the algorithm. It's just the algorithm. What really hurt Apple even
more than that was within days, the next day, Steve Wozniak,
one of the co-founders of Apple, put this Twitter out. "I'm a current Apple
employee and the founder, and the same thing
happened to us, 10 times." Meaning Steve got 10 times
the credit as his wife in the same algorithm. "Some say to blame Goldman
Sachs," et cetera, et cetera. But Apple shares
the responsibility. Not a good rollout. So for Apple Credit
Card, they'll survive. Apple is a big company. They'll probably fix
these biases, but not a particularly good rollout,
as you can well imagine, in their models.

Romain. GUEST SPEAKER: Alida
has her hand up. GARY GENSLER: Please, Alida. AUDIENCE: Yes, you mentioned
cash flow lending earlier. And that really falls under the
merchant cash advance business. There's been a lot of
debate about that being– it's not very
regulated right now, but that becoming
more regulated. Is an entry point like Toast and
other companies that have just launched these new products– [INAUDIBLE] speed
up the regulations around that business? GARY GENSLER: So Romain
or Alida I missed a word. Which company did
you say in there? AUDIENCE: Toast. GARY GENSLER: Oh, Toast, OK. So Toast– for people
to be familiar, Toast started in the
restaurant payment space. And it was basically trying
to provide hardware, tablets. They thought that
the point of sale would be facilitated if
servers had a tablet. And so they were sort of in
the hardware, software space. They found themselves getting
into the payment space very quickly, and
then, of course, data.

And with that data, they
could do a whole cash flow underwriting. And then they started
Toast Capital, where they would make
loans, small business loans, to these restaurants. And I think they have 25,000
or 30,000 restaurants that are in their client list. And they did a round C
funding earlier this year at $4.9 billion. So this is working, of
course until the crisis. And so then the question
is about regulation about cash flow underwriting. I'm not familiar
enough with how Toast feels, even though I've met the
founders and things like that. They're a Boston company. But I think they're dealing
with the same set of regulations that everyone is, which I'm
going to turn to right now in the last eight minutes. But cash flow
underwriting, Alida, are you worried about
something specifically about cash flow underwriting? Because maybe I'll
learn from your concern. AUDIENCE: I [INAUDIBLE]
if you look at– I would consider that to be
like a merchant cash advance product, and those
are not actually considered loans
in the regulations. Now, there's a lot of
movement toward having those products being
considered loans and then fall under different regulatory
standards that [INAUDIBLE]..

GARY GENSLER: So what
Alida is raising– I'm sorry. What Alida raising is, again, in
every political and regulatory process, there is
some definition of what falls within a
regulation, what falls out. You might think, where are
the borders and boundaries of a regulatory environment? What's defined,
really, as a security in the cryptocurrency space? What's defined as an exchange,
an exchange regulation? And hear Alida is saying,
what's defined as a loan? And in Toast's case, doing
cash flow underwriting, it might be considered a cash
advance rather than a loan. So let me do a little
research on that, and we'll come back to that
maybe in a future class when we talk about payments.

But in terms of the consumer
credit law environment, we talked about the Equal
Credit Opportunity Act. The key thing is
not only whether you have disparate treatment,
but whether you have disparate impact. So back to my retreading
analysis, if for some reason you've been reviewing and
your machine algorithms say, aha, all these folks
that are getting retreads should have lower
credit, that might be OK unless you find
that you're treating different protected
classes differently. Are you treating people
of different backgrounds differently, different
genders differently, as Apple certainly was in
Steve Wozniak and his wife? Fair Housing Act, Fair Credit
Reporting Act– the Fair Housing Act has a lot of
these same protected class perspectives.

Fair Credit Reporting– and you
read this in the Mayer Brown piece– the Fair Credit
Reporting Act, you can find yourself as a fintech
company or a data aggregator. The Plaids and the
other data aggregators could find that they were, in
fact, coming under the Fair Credit Reporting Act. They were those vendors
that Cap One might be using. And they, the vendor, might
become Fair Credit Reporting Act companies. And there is usually
a boundary there. Again, this is not a law class,
but these are to highlight. States also have Unfair and
Deceptive Acts and Practices Act. When I was chairman of the
Maryland Consumer Financial Protection Commission, we
went to the state legislature in Maryland. This is a year and a half ago. And said that Maryland's
Unfair and Deceptive Practices Act, UDAP, should be
updated to include abusive. So we sort of broadened
it a little bit. And then privacy laws. This should say general direct– general– I'll correct the
GDPR, but Protection Regulation, and then in the US.

And those are the buckets
that really matter. Sort of trying to close out,
AI, finance, and geopolitics. We've got nearly 200 countries. And those 200 countries,
we're interconnected. We're very
interconnected globally. And we've got a lot
of standards setters, but those standards
setters do not have the authority of lawmakers. So whether it's the
Organization of Economic– OECD, or the guidelines of
things like the securities regulators and the anti-crime
regulators or the banking regulators, these are not
enforceable standards. So what we have is competing
models for AI, finance, and policy. And I note that
because if you're thinking about
starting a company and you operate globally,
sometimes the global law will affect you. GDPR from Europe has
already affected how we deal with privacy in the US. California institutes the
California Consumer Protection Act, it influences
the whole country. So sometimes that
works that way. Romain. GUEST SPEAKER: We have
a question from Akshay. GARY GENSLER: Please, Akshay. AUDIENCE: Hi, professor. So the gender bias
algorithm that you mentioned about the Apple Credit Card,
so the only thing that we can control here is the
data that we're using.

If we are not using
any gender data, and if algorithm turns
out and is creating biases without even using a particular
data which is considered racist or sexist,
so would that be counted as breaking the laws? GARY GENSLER: So here– it's
a great question, Akshay. And again, I caution that
this is not a legal class. But embedded in the US law
and often in other laws is this concept about disparate
treatment and disparate impact. And what you're asking is, what
if you didn't mean to do it? What if there was no
intent, and it's just– wow, that you've extracted this
correlation and all of a sudden there's disparate impact? You could have a
problem in a court. And it's a very
established 50-year-old– there's a lot of
case law around, when would a disparate
impact cause you those headaches and anxiety? And it relates a lot
to explainability. If you can come back
to explainability and you can truly
lay out, this is why, and it has nothing
to do with gender, nothing to do with race, sexual
orientation, and backgrounds, and so forth, a
protected class, you're going to be better
in that court case.

But it would be far better
to have no disparate impact. Then you're in a much
broadly more safe area. AUDIENCE: Got it. Thank you. GARY GENSLER: Romain
other questions? GUEST SPEAKER: Perhaps one
last question from Luke. GARY GENSLER: Oh my God,
Luke, you're always in there. AUDIENCE: I just had a question. GARY GENSLER: Before Luke
goes, is there anybody else? Just, I want to make sure. OK, Luke, you can go. AUDIENCE: It was not a question. It was a comment to
Akshay's question. I'm sure he doesn't
support this. But what he asked is,
isn't it the same thing if a person gives a racist or
a misogynist comment or a hate crime comment? And if they didn't know about
it, is he liable for it? Should a corporate
be held responsible in the same way
a human would be? GARY GENSLER: Well, I
don't know if that's really where Akshay was going, but
I see your point is that basically– and it depends on
the country, Akshay and Luke.

It really does depend
on the country. But here in the US, we'd have
this conceptual framework [? of ?] disparate
treatment, disparate impact. And then explainability
is from another law. But it really should be based
on fairness and inclusion. Everybody's got the same
fair shot, regardless of where the data
comes from at all. So I think that sort of–
we're almost out of time.

test attribution text

Add Comment